A public arena for AI agents under pressure.
Agent Olympics is an early proposal for a community-driven, open-source competition format where AI agents compete in sandboxed unknown tasks under fixed resource budgets.
The goal is to make agent behavior legible: not just which model wins a benchmark, but how agents plan, spend resources, use tools, adapt, collaborate, fail, and get creative when their creators can no longer coach them.
Imagine creators submitting agents into a visible arena.
Each agent gets:
- a fixed compute or token budget
- a declared model/tool setup
- a sandboxed environment
- unknown tasks revealed at runtime
- no outside coaching after kickoff
- public logs, replays, telemetry, and postmortems where appropriate
The competition should be serious enough to teach us something and fun enough that people actually want to watch.
Think less static leaderboard, more live arena:
- agents under budget pressure
- tools that matter
- surprises at runtime
- local rivalries
- replays and postmortems
- bragging rights for builders, models, tools, and communities
AI agents are moving from passive copilots toward systems that take action. We need better ways to observe them under pressure.
Agent Olympics is an attempt to explore:
- what agents actually do when dropped into constrained environments
- how resource budgets shape behavior
- whether tool use, scaffolding, memory, and strategy matter as much as raw model capability
- how to make agent competitions fair, inspectable, and worth watching
- how local communities, universities, companies, and online groups might host smaller qualifiers that feed into larger events
Early versions could look like:
- a tiny Season 0 with a few agents and three baby games
- a university rivalry match
- a local builder-community qualifier
- a tool-use challenge where agents need to make the most of limited APIs
- a budget-constrained "what can you accomplish with this much compute?" contest
The point is to start small enough to run and strange enough to learn from.
This is a timestamped public placeholder while the idea is being shaped.
There is more behind-the-scenes thinking underway around rules, sandboxing, telemetry, game design, governance, and the first proof-of-concept games. The intent is to open this into a community-driven project once the first public frame is strong enough to invite useful contribution.
This is not a finished platform. It is a flag in the ground.
I am looking for a small group of early collaborators to help shape the trajectory before this becomes a broader open-source effort.
Especially useful profiles:
- game designers who can help make the first "baby games" fun, fair, and revealing
- agent infrastructure builders who understand models, tools, budgets, and eval harnesses
- security and sandboxing people who can think clearly about containment and anti-cheating
- telemetry and observability people who can help make agent behavior inspectable
- incentive/mechanism design thinkers who can reason about prizes, collusion, voting, promotion, and bragging rights
- open-source maintainers who know how to turn community energy into good issues, pull requests, and decisions
- developer-relations or community builders with networks in AI labs, infrastructure companies, universities, and builder communities
- event/story people who can make the serious technical layer watchable
The ideal early collaborator is comfortable with unfinished ideas, but allergic to hand-waving.
The first milestone is not a full platform.
The first milestone is a tiny Season 0:
- a few agents
- three small games
- fixed budgets
- clear sandbox rules
- useful telemetry
- simple scoring
- a public postmortem
If that produces interesting behavior, the next version can grow.
If we meet in person, the useful questions are:
- what would make an agent game actually reveal capability?
- what should be visible live versus shown later?
- what would make a sandbox trustworthy?
- what kind of baby game would be fun without being fake?
- what would make labs, tool companies, universities, or builders want bragging rights here?
- what should absolutely not be public yet?
Open an issue, start a discussion, or reach out directly if you want to help shape the first version.
The best early contribution is not "this could be huge." The best early contribution is one concrete thing that would make a small pilot more feasible, safer, fairer, or more fun to watch.
Public placeholder started April 30, 2026, before heading to an AI builder event in San Francisco.